building an environmental sustainability dictionary for ... · in. in dictionary based text mining...
TRANSCRIPT
Building an Environmental Sustainability Dictionary for the IT Industry
Qi Deng
Carleton University
Michael Hine
Carleton University
Shaobo Ji
Carleton University
Sujit Sur
Carleton University
Abstract Content analysis is a commonly utilized
methodology in corporate sustainability research.
However, because most corporate sustainability
research using content analysis is based on human
coding, the research capability and the scope of the
research design has limitations. The relatively recent
text mining technique addresses some of the limitations
of manual content analysis but its usage is often
dependent upon the development of a domain specific
dictionary. This paper develops an environmental
sustainability dictionary in the context of corporate
sustainability reports for the IT industry. In support of
building said dictionary, we develop a standardized
dictionary building process model that can be applied
across many domains.
Keywords: dictionary building process,
environmental sustainability, text mining, IT industry
1. Introduction
Research on corporate sustainability (CS) reporting
has a long history of using a manual content analysis
(MCA) method based on human coding [1-3]. This
choice of methodology is based on many reasons. First,
MCA is a well-established research method and a set of
research procedures that has been developed and
validated to guide the research process. Second, MCA
has been widely used as a way to make valid inferences
from textual data, which happens to be the main content
format of corporate sustainability disclosures. Third,
MCA is an alternative method to examine issues, which
would be time and resource intensive and too obtrusive
if studied using other techniques, e.g., direct
observations. For researching corporate sustainability
reporting, content analysis of what companies have
disclosed regarding their sustainability performances
might be the most effective and appropriate
methodology [see 4-12]. However, with the large
volumes and high velocity of digitized textual materials
on corporate sustainability reports increasingly being
made available, MCA becomes constrained and less
efficient as it is extremely time consuming and prone to human error. Increasingly, MCA faces criticism about
coding reliability and potential coding errors caused by
coder fatigue, misapplication of the coding schema, and
potential disagreement between coders on particular
attribute values [13-15].
Text mining (or automated content analysis, ACA)
has the potential to address these problems. Text mining
refers to the process of detecting patterns or knowledge
from unstructured or semi-structured text and it has
many advantages over MCA, such as enhanced
reliability, elimination of manual coding errors, low cost,
and the capability for analyzing large amounts of data in
considerably short time period [4, 16-18].
In many cases, text mining is reliant on a thesaurus-
like dictionary. A typical dictionary includes categories
that contain words, word stems, and phrases. The
frequencies of the words, stems, phrases, and thus
categories, are counted and, based on these frequencies,
the relative importance or changes over time of the
central concepts in the texts can be determined. The
dictionary allows researchers to systematically assess
different aspects of the core concept they are interested
in. In dictionary based text mining efforts, the quality
of the results is largely dependent on the quality of the
dictionary. Developing a dictionary is an iterative and
time-consuming process which could last from months
to years [17, 19]. However, once developed, a dictionary
can be applied to any text mining projects related to the
same domain and is very useful for document indexing
and categorization as well as document retrieval [19].
There is no doubt that research on corporate
sustainability reporting can benefit from the text mining
method, especially in view of the increased digital
availability of large volumes of CS related textual data.
While some previous corporate sustainability
research has applied text mining method, most efforts
have been at an introductory level and no related
dictionary has been developed [see 4, 15, 20].
Considering its potential capability and current wide
adoption in other research areas (e.g., Tourism,
Agriculture, Political Science, Medical Science, and
Psychology), the most possible reason for this under-
utilization is the lack of a valid dictionary. Thus, to
facilitate future proliferation of text mining in corporate
sustainability research the necessary first step is to
develop a useful and valid dictionary. While several
papers using text mining touch upon the problem of
building dictionary, in most, if not all, of them, the
processes used to build the dictionary are not described
clearly and are more or less subjective. To the best of
950
Proceedings of the 50th Hawaii International Conference on System Sciences | 2017
URI: http://hdl.handle.net/10125/41264ISBN: 978-0-9981331-0-2CC-BY-NC-ND
our knowledge, no systematic dictionary building
process has been documented in any manuscript.
This paper has two main objectives: 1) to develop a
general dictionary building process model; 2) to
actualize the aforementioned process in building a
dictionary for detecting environmental sustainability
topics for IT companies and demonstrate the initial
dictionary’s usage. The rest of the paper is organized as
follows. Section 2 presents a dictionary building process
model developed based on a review of previous related
research. The method used to build the environmental
sustainability dictionary and the result are described in
Section 3. Section 4 presents a demonstration of using
the dictionary to analyze the newspaper articles. Section
5 presents the discussion and conclusion.
2. Dictionary Building Process
To build a dictionary, one needs to identify the “right”
words and/or phrases in the corpus and assign them into
different categories that represent concepts that the
researcher is interested in. For example, to build an
emotion dictionary which can be used to analyze online
product comments, researchers probably identify the
words “satisfy”, “good”, and “useful” as representative
of positive emotion and the words “terrible”, “angry”,
“useless” as that of negative emotion.
Previous literature on dictionary building has two
streams: automatic dictionary building and semi-
automatic dictionary building. Rooted in information
extraction research, automatic dictionary building
usually involves extracting key words and/or phrases
automatically based on learning algorithms and
evaluating the resulting dictionary by experiments or
comparing it with existing dictionaries [see 21-26]. In
semi-automatic dictionary building researchers make
their own dictionary inclusion judgements on words
and/or phrases with the assistance of text analysis
software.
In this paper, we develop a process model to support
dictionary building consistent with the existing semi-
automatic dictionary building research. A preliminary
search of literature on text mining and addressing
dictionary building resulted in 15 papers. None of these
papers adopted a standardized dictionary building
process as is common in automatic dictionary building
research. Following an inductive approach, where
possible, we: analyzed the descriptions of dictionary
building processes (or lack thereof) in these papers,
summarized the steps adopted (see Table 1), and
subsequently derived a general dictionary building
process.
Table 1. Summary of dictionary building process
Citation Dictionary Corpus
Creation
Pre-processin
g
Entry
Identification &
Categoriz
ation
Extension & Simplification
Validat
ion
Synonym
s &
Antonyms
Stemmin
g
Weight
ing
[27] Online image and video
subject X X
[28] Job description X X
[29] Tone in financial text X X
[30] Corporate philanthropy X X
[31]
External validation, shareholder alignment,
market performance and
accounting performance
X X X
[32] Rational and normative
words X X X
[33] Precautionary principle X X X
[34] Forest value X X X
[35] Auditing research topics X X X
[36] Danish Adverse Drug
Events X X X X
[37]
Competency-related terms
in business intelligence and
big data job ads.
X X X X
[38] Privacy related issues X X X X X
[39] Privacy related issues X X X X X
[40] Policy agendas X X X X
[18] Public leadership image X X X
951
We name the resulting documentation the “semi-
automatic dictionary building process (S-DBP)”. The S-
DBP includes five steps, namely, corpus creation, pre-
processing, word and phrase (entries) identification and
categorization, extension and simplification, and
validation. While iteration within the steps is common
we will discuss the steps in linear fashion.
Step 1. Corpus creation. The corpus is the source
documents from which the dictionary is developed. It
usually consists of multiple documents which include
rich textual contents related to the topic of the dictionary.
Creating a corpus involves selecting the right textual
sources for future processing. Since the dictionary is
derived from the corpus, its quality is directly dependent
on the documents in the corpus.
Previous studies have not generally addressed the
assessment of corpus. Three features of the corpus could
be considered to decide whether the corpus is
“adequate”. First, the corpus should be relevant. It
should include the contents which are consistent with
the theme of the dictionary to be built. Second, the
corpus should be appropriate. Since the subsequent
steps are mainly based on the analysis of text, the
original corpus should include mainly textual contents,
instead of numeric or pictorial contents. Third, the
corpus should be complete. For example, in order to
build a dictionary of forest values, Bengston and Xu [34]
created a corpus which includes articles by forest
economists, traditional foresters, forest ecologists,
landscape architects, aestheticians, environmental
philosophers, environmental psychologists, Native
Americans, and so on. To be complete does not mean
that the corpus should include every related document,
rather its should ensure that the richness and
completeness of the corpus should be adequate to
support the dictionary building. The criterion of
“completeness” is especially important for the process
of a building dictionary with pre-specified categories. If
the corpus does not cover all pre-specified categories,
neither will the dictionary.
Step 2. Pre-processing. The aim of this step is to
prepare the corpus for further analysis using data
cleaning techniques including: stop words removal [see
37], unnecessary information removal [see 35-36],
reducing phrases to single words [see 38-39], spelling
correction, and so on. Usually the pre-processing is
conducted with the help of text analysis or text mining
software. Currently, there are many computer-aided text
analysis (CATA) software can assist with the pre-
processing step, such as WordStat and RapidMiner
among others. Whether to conduct this step and which
techniques to be used are decisions which are made by
researchers based on the requirement of the dictionary.
Of the 15 identified papers, 5 include this step and 10 do
not conduct this step.
Step 3. Entry identification and categorization.
Usually, a dictionary includes three basic elements: the
entries (words, word stems and phrases), the categories,
and the association between the entries and the
categories. Categories, according to Weber [41, p. 140]
are “a group of words [and phrases] with similar
meaning and/or connotations”. In this step, researchers,
who are familiar with the theme of the dictionary,
examine each entry in the list developed in the second
step and decide whether the entry should be retained and
into which category the entry should be assigned. Entry
identification and categorization are typically carried
out by researchers with assistance of text analysis
software. Many projects do not have pre-specified
categories and are more exploratory in nature. In these
situations, dictionary categories are derived from the
content of the corpus itself. Typically, this is done with
the aid of a ‘topic extraction’ feature within text mining
software that aids in uncovering thematic structure of
the processed text. Topic extraction is usually
implemented using latent semantic analysis or latent
dirichlet allocation.
Researchers often determine cut-off criteria and
exclude entries from the dictionary that do not meet the
criteria. Popular cut-off criteria include term frequency,
and frequency of the documents in which one entry
occurs. For example, “terms occurring in less than 1%
of the documents” was used in Lesage & Wechtler [35]
and Debortoli, Müller & vom Brocke [37] as cut-off
criterion, while “terms occurring more than 30 times”
was used in Abrahamson & Eisenman [32] as cut-off
criterion. TF*IDF is another popular cut-off criterion.
TF refers to term frequency and IDF refers to inverse
document frequency. Although TF*IDF has not been
used in the papers we reviewed, it is a standard way of
culling words up front. The usage of this metric is based
on the assumption that the more frequent a term occurs
in a document, the more representative it is of the
document’s content yet, the more documents in which
the term occurs, the less important the term is in
distinguishing different documents’ content from each
other. So, if the purpose of the research is to distinguish
between documents, as it is in classification tasks,
TF*IDF is extremely important.
As our review indicates, the cut-off criterion is
usually an arbitrary decision made by researchers based
on the scope of the corpus or a decision to follow
established criteria levels from previous studies. In most
of the studies we reviewed, the entry identification and
categorization are conducted by single researcher.
However, it can be performed by multiple researchers as
well. In the multi-coder case, the concept of inter-coder
reliability is introduced as an assessment of the word
categorization [see 32]. The result of this step is an
952
initial dictionary which could be further modified or
directly applied to analyze additional text documents.
Step 4. Extension and simplification. Many
techniques can execute this task, but generally speaking,
the most common ones are synonym and antonym
extension, stemming, lemmatization and weighting. The
synonyms and antonyms extension means to add
synonyms (and antonyms) for the initial words to the
dictionary. Because of the various wording preference,
different terms might be used by different authors to
express the same meaning. To extend the dictionary by
including synonyms (and antonyms) can, in some
degree, increase the generalizability of the dictionary.
The entries in the dictionary are not necessarily whole
words or phrases, but are often reduced by stemming or
lemmatizing. Stemming is a more rudimentary approach
where words are simply truncated. For example, the
word “having” maybe stemmed to “hav*”.
Alternatively, lemmatizing aims to retain the
morphology of the word and would thus reduce “having”
to “have”. The choice of approach is project dependent.
Stemmers are faster and simpler but lemmatization is
more accurate. In this way, the dictionary can be
simplified without costing the accuracy and
effectiveness. Weighting means to weight terms based
on their occurrence in and across documents. It is
usually performed by applying the commonly used
TF*IDF (Term Frequency-Inverse Document
Frequency) weighting scheme. Compared with
synonyms and antonyms extension, stemming, and
lemmatization weighting is less commonly used. But, in
some special cases this technique can promote the
occurrence of rare terms and discounts the occurrence of
more common terms [37, 42].
Step 5. Validation. The fourth step results in an
extended and simplified dictionary that should be
validated before being widely applied. Of the 15 studies
reviewed, 9 report some form of validation of the
dictionary. As the review shows, the most common
validation method is to examine the key-word-in-
context (KWIC), following by to compare-with-human-
coding (CWHC), demonstration, and expert validation.
Since the same entry might have different meaning in
different context, it is necessary to have a look at the
actual usage of the entry in the corpus to determine
whether the entry is the accurate indicator of the concept
the researcher perceives it to indicate. Another
validation method is to compare the automated coding
results with human coding results. The similarity
between the automated coding results and human coding
results are the primary indicator of the validity of the
dictionary. Researchers also can validate the dictionary
by demonstration (to actually apply the dictionary) or by
expert validation (to have an expert on the theme of the
dictionary to have a review of the dictionary).
One item of note is that the S-DBP aims to provide
instructional guidelines, rather than impose
requirements, for researchers interested in domain-
specific dictionary building. Although we illustrate the
dictionary building process as a sequential step-by-step
process, in reality dictionary building is an iterative
process where steps are often revisited. For example, if
the quality and quantity of the entries identified in step
3 are below one’s expectation, one might need to re-
think about the corpus creation. After validation, one
might need to re-think the whole process to see if there
are any improvements one can do to make the dictionary
better one. To build a comprehensive dictionary is a
long-term activity which could last from months to
years [19, 40, 43]. However, not every dictionary is
necessarily comprehensive. The scope of the dictionary
is decided based on the purpose of the research. The
dictionary can be used confidently as long as it is
comprehensive enough to support its purpose. In next
section, we describe the process of building an
environmental sustainability dictionary for IT
companies following the S-DBP approach.
3. Environmental Sustainability Dictionary
We follow the S-DBP described above to build a
dictionary for environmental sustainability of IT
companies. With the rise of the concept of “Green IT”,
the IT industry has paid increasing attention to
environmental sustainability. We use WordStat from
Provalis Research to support the dictionary building
process. WordStat has been used extensively in text
analysis related research.
Step 1: Corpus creation. Corporate sustainability
reports of IT companies from the 2015 Fortune 500
were collected and used to create the corpus for
dictionary building for three reasons. First, corporate
sustainability reports usually include economic, social,
and environmental sustainability performance content;
it is thus related. Second, despite the presence of some
numerical data, most of the contents in the corporate
sustainability report are textual data, and therefore
appropriate. Third, the corporate sustainability report is
one of the most important artefacts to communicate a
company’s sustainability performance to its
stakeholders. Therefore, it generally includes every
aspect of the company’s sustainability performance and
can be considered complete. Of the 49 IT companies
included in the 2015 Fortune 500, 28 issued annual
corporate sustainability reports, 10 issued online
sustainability disclosures, and 11 did not disclose
corporate sustainability information. To improve the
corpus’ relatedness, we only collect the environmental
section from the CS reports and online disclosures from
953
2015. This results in 751 pages (reduced from 2,119
pages) of CS report contents and 53 pages of online
disclosure contents. In total, the initial corpus consists
of 38 documents (reports or online disclosures), which
include 804 pages of environmental sustainability
related contents.
Step 2: Pre-processing. After importing the initial
corpus into WordStat, we conducted two steps of pre-
processing; spelling check and stop word (e.g., “a”,
“and”, “or”, etc.) removal. Although corporate
sustainability reports and online disclosures are official
publications and, usually, they do not include spelling
mistakes, it is still necessary to conduct a spelling check
before further analysis because the format of the textual
data might change during the data importing step. For
example, the original phrase, “environmental
sustainability”, might become
“environnmentalsustainability” after being imported.
Since these format changes influence the frequency
analysis later, it is necessary to deal with them before
conducting next step. The spelling check can be
conducted with the help of built-in functions of
WordStat.
WordStat has a built-in stop word dictionary which
includes common stop words and can be refined by
researchers according to the research objective.
Enabling the stop word removal function will
automatically exclude the stop words from the
subsequent text analysis. We used the default stop words
dictionary because it does not include sustainability-
related words, thus, will not impact the text analysis
later.
Step 3: Entry Identification & Categorization. In this
paper, we adopted an iterative process to identify and
categorize the environmental sustainability-related
entries. The 38 documents were randomly divided into
a training set and a testing set, with each including 19
documents. We then developed an initial dictionary
from the training set. We then refined the initial
dictionary by applying it in the testing set to see whether
there are qualified entries in leftover entry set. The
testing set was randomly divided into four subsets (5
documents for three subsets and 4 for one subset) and
the initial dictionary was refined through four rounds.
Both the initial development and the later refinement
followed similar entry identification and categorization
process as described below.
Entry categorization. We adapted the environmental
sustainability categories of the GRI G4 reporting
framework to support the entry categorization. This
approach is consistent with many corporate
sustainability studies [see 3, 5, 7, 15]. The GRI G4
environmental sustainability framework covers twelve
related aspects including: materials, energy, emissions,
water, biodiversity, effluents & waste, products &
services, compliance, transport, supplier environmental
assessment, environmental grievance mechanisms, and
overall. We remove “overall” from our categorization
framework because it is fully overlapped with other
categories. Therefore, we pre-specified eleven
categories.
Entry identification. In the initial development stage,
and after pre-processing, the 19 documents in training
set contained 7,487 words. After applying the cut-off
criterion of “occurring in no less than 2 documents”,
3,865 words are retained. After applying the cut-off
criterion of “occurring no less than 5 times with max
words of 4”, 915 phrases were generated. The first
author then manually reviewed the 4,780 entries (both
words and phrases) and identified environmental
sustainability-related entries which represented the
eleven categories of the coding schema described above.
Each identified entry was categorized based on the
examination of keywords in context (KWIC). The initial
attempt resulted in a dictionary containing 261 entries.
We then applied the dictionary in the testing subsets and
examined the leftover words following the same cut-off
criteria to see whether there were additional qualified
entries. After four rounds of refinement, the dictionary
included 287 entries.
Step 4: Extension & Simplification. For the words in
the initial dictionary, we examined their synonyms,
which also occur in the documents, to see whether they
should be included in the dictionary. Similar to the
initial coding, this step was also guided by the coding
schema and with the help of KWIC. This step generated
15 new words. We did not conduct stemming or
lemmatization here because we found that, sometimes,
the different tenses of one word had different meanings.
Finally, since this was the first step to build an
environmental sustainability dictionary, we did not
weight the entries either.
Step 5: Validation. We conducted two rounds of
validation of the dictionary. In the first round, we
designed a task of re-coding the previously identified
entries into the dictionary categories. A PhD student,
who was familiar with corporate sustainability concepts,
was hired to conduct this task. The task included two
rounds. In the first round, the student was asked to re-
categorize the entries in the dictionary into the eleven
categories based on our coding schema without the
assistance of the KWIC capability. In the second round,
the student was asked to perform the task with the help
of the KWIC. In both rounds, the student did not know
the original coding results of the entries. The reliability
between original coding and additional coding is shown
in table 2 below.
954
Table 2. Inter-Coder reliability of the dictionary validation
No. Category Number of Entries Reliability
Round 1 Round 2
1 BIODIVERSITY 13 54% 62%
2 COMPLIANCE 23 100% 100%
3 EFFLUENTS & WASTE 45 53% 80%
4 EMISSIONS 38 82% 100%
5 ENERGY 73 75% 93%
6 ENVIRONMENTAL GRIEVANCE MECHANISMS 3 67% 67%
7 MATERIALS 27 48% 70%
8 PRODUCTS & SERVICES 24 63% 58%
9 SUPPLIER ENVIRONMENTAL ASSESSMENT 16 94% 88%
10 TRANSPORT 24 79% 79%
11 WATER 16 94% 100%
All Entries 302 72% 85%
*Scale of the inter-coder reliability: 0.21-0.40 (Fair); 0.41-0.60 (Moderate); 0.61-0.80 (Substantial); 0.81-1.00
(Almost Perfect) [44-45].
As shown in Table 2, the interrater reliability
improved from ‘substantial’ (72%) to ‘almost perfect’
(85%) with the help of KWIC. The first author re-
examined every entry coded differently from the second
coder and discussed the entry context with the second
coder. The dictionary was then refined based on the
discussion. The final dictionary included 302 words and
phrases, a portion of which are shown in Table 3. One
thing to notice in this dictionary is that the entries are
not equally distributed in different categories. The
variety of the distribution reflects that the IT companies
pay different attention to different environmental
sustainability aspects. For example, it is clearly shown
in Table 3 that IT companies have paid more attention
to Energy, Emission, and Effluent & Waste than
Biodiversity, Water, and Environmental Grievance
Mechanisms. The demonstration of the generated
dictionary is presented in the next section.
Table 3. Dictionary of environmental sustainability for IT industry (sample)
No. Category Entries
1 BIODIVERSITY biodiversity; conservation; plants; tree; wildlife; …
2 COMPLIANCE compliance; compliant; law; regulation; …
3 EFFLUENTS &
WASTE
nonhazardous; composted; disposal; electronic waste; ewaste; landfill; product
end of life; recycling; waste; remanufacturing; reuse; …
4 EMISSIONS carbon offset; greenhouse; air emissions; air pollution; carbon; carbon dioxide;
carbon neutral; dioxide; emission; footprint; …
5 ENERGY
air conditioning; biogas; cells; clean energy; cooling; electricity; energy;
energy star; fuel; gas; gasoline; grid; heating; hydro; kilowatt; lamps; led;
lighting; power; renewable energy; solar; wind; wind farm; …
… … … … … … …
4. Demonstration
The purpose of the demonstration is to show how the
resulting dictionary can be used in an analysis of
environmental sustainability for technology companies.
Because of the small amount of data being analyzed and
given the nascent stages of dictionary development we
are cautious about drawing any conclusions from the
results reported below. At this stage, we consider the
demonstration as a “proof of concept” only.
For the demonstration, we collected environmental
sustainability related newspaper articles from
LexisNexis. To limit the scope for ease of demonstration,
we only search related articles published in New York
955
Times from 2001 to 2015, which covers the 15 years
during which the corporate sustainability achieved a
rapid awareness worldwide. The method we used to
search the articles is as follows:
HLEAD (Corporate Name, e.g., Apple,
Microsoft, etc.) AND BODY (social
responsibility) OR BODY (corporate
responsibility) OR BODY (corporate citizenship)
OR BODY (sustainability) OR BODY
(environmental)
In total, the search results in 698 articles. Under
some corporate names (i.e., Apple, Amazon), the search
tends to result in more unrelated articles. The reason is
that these searches result in some articles that are
actually about apple, the fruit, and Amazon, the forest.
We thus reviewed the first paragraph of each article to
make sure that we only include sustainability related
articles. This resulted in 449 articles. An import
template was designed and the articles were brought into
QDAMiner / WordStat for future analysis.
Using WordStat we detected all the words/phrases
from the dictionary in the articles and generated a
contingency table showing the percentage of words in
each of the dictionary categories across year of
publication. This data can then form the basis of
analysis that adds insight into how the different topics
(represented by categories) of environmental
sustainability ebb and flow across time as reported by a
media source. Because the outcome of the application
of text mining is often a contingency table, it is typical
to report results using correspondence analysis (CA).
CA is a method that allows the graphical representation
of contingency table data in low dimensional space [46].
CA has been successfully used in a variety of domains
including marketing [47], tourism management [48-50],
teaching and learning [51] among others. While there are several types of CA maps available,
Greenacre states that “the symmetric map is the best
default map to use” (46: 267). The symmetric map
typically provides a ‘nicer-looking’ representation than
the asymmetric approach which often compresses the
primary coordinates of the row profiles towards the
centre of the map to allow the display of the extreme
vertices of the column profiles (essentially creating a
map that is more difficult to visualize than a symmetric
map). The CA map of the contents of New York Times
articles as detected by the sustainability dictionary is
shown in Figure 1 below.
Figure 1. CA Map of environmental sustainability topics from the media across time
956
The point at which the axes cross represents the
average yearly profile of environmental sustainability
topics. Note that some years are not represented in
the map as there was insufficient data to detect an
appropriate amount of relevant words/phrases. If we
look primarily at the horizontal axis, which in CA
explains more of the variance than the vertical axis,
we see that the yearly profiles are the most different
between {2002; 2010; 2013} and 2011as the
horizontal distance between these years is the greatest.
The {2002; 2010; 2013} profiles are fairly similar but
distinguished by proportionally more entries in water
in 2013 and proportionally more entries on energy
and ‘products and services’ in 2002 and 2010. The
profiles of 2014 and 2006 articles are similar and
proportionally contain more content related to
transportation than do other years’ articles. Finally,
the 2011 profile is the most unique of the reported
years with proportionally more entries dealing with
‘effluent and waste’ and ‘supplier environmental
assessment’.
5. Discussion and Conclusion
Corporate sustainability research can benefit from
adopting a text mining approach. To promote and
maximize the benefit, building a dictionary is a
necessary first step. The aim of this paper is to take
the first step. This paper has two major contributions.
First, although research on dictionary building
already exists, as far as we know, none of them follow
or propose a standardized dictionary building process.
Based on previous automated content analysis studies,
we propose a semi-automatic dictionary building
process, which includes five steps, namely, corpus
creation, pre-processing, words identification and
tagging, extension and simplification, and validation.
Notably, the dictionary building is an iterative
process which could last from months to years.
Second, we have built an initial environmental
sustainability dictionary for the IT industry. Although
this dictionary is only an initial version and still need
further modifications, we do believe that the
development of such dictionary will promote the
adoption of text mining method in corporate
sustainability area and, in turn, facilitate the research
in this area.
This paper is not without limitations. First, the
corpus created for dictionary building is limited. We
only included the most recent corporate sustainability
reports (and online disclosures) in the corpus.
Although, logically, the CS reports should cover
every aspects of the companies’ environmental
sustainability activities, the dictionary building
should probably incorporate data from different
sources to ensure completeness. Despite of the data
being sourced from company reports, future research
could also incorporate data from mainstream media,
non-profit organizations, government, among others.
Second, during the dictionary building process, we
made some arbitrary decisions. For example, we use
the cut-off criterion of “occurring no less than 2
documents” without evaluating the impacts of the
criterion on our results. As far as we know, previous
research has not addressed the impacts of such cut-
off criteria on the dictionary building results.
However, for dictionary building research, such
evaluation is significant. Future research could
investigate that area. Third, due to the limitation of
time and scope, we only included one extra coder to
validate the dictionary. The increase of inter-coder
reliability is not without an experience threat. Future
research should include multiple coders and multiple
trials to validate the dictionary. After development,
the dictionary needs more robust validation. Since the
quality of the text mining research is limited by the
quality of the dictionary used, it is necessary and
important to make sure the dictionary is adequate. To
our knowledge, there is limited research addressing
what might constitute an adequate dictionary [41].
We call for future studies to investigate this issue
further.
In conclusion, the objective of the proposed S-
DBP is to provide researchers interested in dictionary
building with a general guideline to follow. Our hope
is that the S-DBP could provide a basic model for
future dictionary building. The second contribution
of this study is the development of a dictionary for
studying environmental sustainability in the IT
industry. To our knowledge, it is the first dictionary
developed for the corporate sustainability field.
6. References
[1] J.A. Arevalo, “Critical reflective organizations: An
empirical observation of global active citizenship and
green politics”, Journal of Business Ethics, 96(2), 2010, pp. 299-316.
[2] M. Asif, C. Searcy, P.D. Santos, and D. Kensah, “A
review of Dutch corporate sustainable development reports”, Corporate Social Responsibility and
Environmental Management, 20(6), 2013, pp. 321-
339. [3] Delai, and S. Takahashi, “Corporate sustainability in
emerging markets: Insights from the practices
reported by the Brazilian retailers”, Journal of
Cleaner Production, 47, 2013, pp. 211-221. [4] R. Barkemeyer, L. Preuss, and L. Lee, “On the
effectiveness of private transnational governance
regimes – evaluating corporate sustainability reporting according to the global reporting
957
initiative”, Journal of World Business, 50(2), 2015,
pp. 312-325. [5] M.J. Bonilla-Priego, X. Font, and M. del Rosario
Pacheco-Olivares, “Corporate sustainability reporting
index and baseline data for the cruise industry”, Tourism Management, 44, 2014, pp. 149-160.
[6] J.M. Comas Martí, and R.W. Seifert, “Assessing the
comprehensiveness of supply chain environmental
strategies”, Business Strategy and the Environment, 22(5), 2013, pp. 339-356.
[7] D. de Grosbois, “Corporate social responsibility
reporting in the cruise tourism industry: A performance evaluation using a new institutional
theory based model”, Journal of Sustainable
Tourism, 2015, pp. 1-25. [8] C. Deegan, M. Rankin, and J. Tobin, “An
examination of the corporate social and
environmental disclosures of BHP from 1983-1997:
A test of legitimacy theory”, Accounting, Auditing & Accountability Journal, 15(3), 2002, pp. 312-343.
[9] A. Fonseca, “How credible are mining corporations’
sustainability reports? A critical analysis of external assurance under the requirements of the international
council on mining and metals”, Corporate Social
Responsibility and Environmental Management, 17(6), 2010, pp. 355-370.
[10] S.K. Fuisz-Kehrbach, “A three-dimensional
framework to explore corporate sustainability
activities in the mining industry: Current status and challenges ahead”, Resources Policy, 46, 2015, pp.
101-115.
[11] L. Gatti, and P. Seele, “Evidence for the prevalence of the sustainability concept in European corporate
responsibility reporting”, Sustainability Science,
9(1), 2014, 89-102. [12] R. Hahn, and R. Lülfs, “Legitimizing negative
aspects in GRI-oriented sustainability reporting: A
qualitative analysis of corporate disclosure
strategies”, Journal of Business Ethics, 123(3), 2014, pp. 401-420.
[13] W.J. Potter, and D. Levine-Donnerstein, “Rethinking
validity and reliability in content analysis”, Journal of Applied Communication Research, 27(3), 1999,
pp. 258-283.
[14] N. Scott, and A.E. Smith, “Use of automated content analysis techniques for event image assessment”,
Tourism Recreation Research, 30(2), 2005, pp. 87-
91.
[15] D.L. Gill, S.J. Dickinson, and A. Scharl, “Communicating sustainability: A web content
analysis of North American, Asian and European
firms”, Journal of Communication Management, 12(3), 2008, pp. 243-262.
[16] R.E. Orwig, H. Chen, and J.F. Nunamaker, “A
Graphical, Self-Organizing Approach to Classifying Electronic Meeting Output”, Journal of the American
Society for Information Science, 48(2), 1997, pp.
157-170. [17] R. Morris, “Computerized content analysis in
management research: A demonstration of
advantages & limitations”, Journal of Management,
20(4), 1994, pp. 903-931.
[18] L. Aaldering, and R. Vliegenthart, “Political leaders
and the media. Can we measure political leadership images in newspapers using computer-assisted
content analysis?”, Quality & Quantity, 2015, pp. 1-
35. [19] N. Péladeau, and C. Stovall, “Application of Provalis
Research Corp.’s statistical content analysis text
mining to airline safety reports”, Retrieved from
http://flightsafety.org/files/Provalis text mining report.pdf, 2005.
[20] N. Reuter, S. Vakulenko, J. vom Brocke, S.
Debortoli, and O. Müller, “Identifying the role of information systems in achieving energy-related
environmental sustainability using text mining”,
Proceedings of the 22nd European Conference on Information Systems, Tel Aviv, Israel, track 22, 2014,
article 12.
[21] J.T. Chang, H. Schütze, and R.B. Altman, “Creating
an online dictionary of abbreviations from MEDLINE”, Journal of the American Medical
Informatics Association, 9(6), 2002, pp. 612-620.
[22] E. Riloff, “Automatically constructing a dictionary for information extraction tasks”, Proceedings of the
11th National Conference on Artificial Intelligence,
Washington, D.C., 1993 July 11-15, pp. 811-816. [23] Y. Takayama, Y. Tomiura, K.R. Fleischmann, A.S.
Cheng, D.W. Oard, and E. Ishita, “Automatic
dictionary extraction and content analysis associated
with human values”, Information Engineering Express, 1(4), 2015, pp. 107-118.
[24] N. Oostdijk, S. Verberne, and C.H. Koster,
“Constructing a broad-coverage lexicon for text mining in the patent domain”, Proceedings of the 7th
International conference on Language Resources
and Evaluation, Valletta, Malta, 2010 May, pp. 17-23.
[25] M.Z. Asghar, A. Khan, S. Ahmad, and B. Ahmad,
“Subjectivity lexicon construction for mining drug
reviews”, Science International, 26(1), 2014, 145. [26] E. Riloff, and W. Lehnert, “Information extraction as
a basis for high-precision text classification”, ACM
Transactions on Information Systems (TOIS), 12(3), 1994, pp. 296-333.
[27] J.R. Smith, and S.F. Chang, “Searching for images
and videos on the world-wide web”, IEEE Multimedia Magazine, Retrieved from
http://www.ee.columbia.edu/ln/dvmm/publications/9
6/smith96e.pdf, 1996.
[28] J.R. Park, C. Lu, and L. Marion, “Cataloging professionals in the digital environment: A content
analysis of job descriptions”, Journal of the
American Society for Information Science and Technology, 60(4), 2009, pp. 844-857.
[29] T., Loughran, and B. McDonald, “When is a liability
not a liability? Textual analysis, dictionaries, and
10‐Ks”, The Journal of Finance, 66(1), 2011, pp.
35-65. [30] B. de-Miguel-Molina, V. Chirivella-González, and B.
García-Ortega, “Corporate philanthropy and
community involvement. Analysing companies from
France, Germany, the Netherlands and Spain”, Quality & Quantity, 2015, pp. 1-26.
958
[31] J.B. Wade, J.F. Porac, and T.G. Pollock, “Worth,
words, and the justification of executive pay”, Journal of Organizational Behavior, 18(1), 1997, pp.
641-664.
[32] E. Abrahamson, and M. Eisenman, “Employee-management techniques: Transient fads or trending
fashions?”, Administrative Science Quarterly, 53(4),
2008, pp. 719-744.
[33] A. Kirilenko, S. Stepchenkova, R. Romsdahl, and K. Mattis, “Computer-assisted analysis of public
discourse: A case study of the precautionary
principle in the US and UK press”, Quality & Quantity, 46(2), 2012, pp. 501-522.
[34] D.N. Bengston, and Z. Xu, “Changing national forest
values: A content analysis”, Retrieved from http://www.nrs.fs.fed.us/pubs/rp/rp_nc323.pdf. 1995.
[35] C. Lesage, and H. Wechtler, “An inductive typology
of auditing research”, Contemporary Accounting
Research, 29(2), 2012, pp. 487-504. [36] R. Eriksson, P.B. Jensen, S. Frankild, L.J. Jensen,
and S. Brunak, “Dictionary construction and
identification of possible adverse drug events in Danish clinical narrative text”, Journal of the
American Medical Informatics Association, 20(5),
2013, pp. 947-953. [37] S. Debortoli, O. Müller, and J. vom Brocke,
“Comparing business intelligence and big data
skills”, Business & Information Systems
Engineering, 6(5), 2014, pp. 289-300. [38] A.J. Gill, A. Vasalou, C. Papoutsi, and A.N. Joinson,
“Privacy dictionary: A linguistic taxonomy of
privacy for content analysis”, Proceedings of the SIGCHI Conference on Human Factors in
Computing Systems, ACM, 2011 May, pp. 3227-
3236. [39] A. Vasalou, A.J. Gill, F. Mazanderani, C. Papoutsi,
and A. Joinson, “Privacy dictionary: A new resource
for the automated content analysis of privacy”,
Journal of the American Society for Information Science and Technology, 62(11), 2011, pp. 2095-
2105.
[40] Q. Albaugh, J. Sevenans, S. Soroka, and P.J. Loewen, “The automated coding of policy agendas:
A dictionary-based approach”, Proceedings of the
6th Annual Comparative Agendas Conference, Atnwerp, Beligum, 2013 June, pp. 27-29.
[41] R.P. Weber, “Measurement models for content
analysis”, Quality & Quantity, 17(2), 1983, pp. 127-
149. [42] C.D. Manning, P. Raghavan, and H. Schütze,
Introduction to information retrieval, Cambridge
University Press, Cambridge, 2008. [43] J.W. Pennebaker, R.L. Boyd, K. Jordan, and K.
Blackburn, “The Development and Psychometric
Properties of LIWC2015. UT Faculty/Researcher Works”, Retrieved from
http://hdl.handle.net/2152/31333, 2015.
[44] J. Cohen, “A coefficient of agreement for nominal scales”, Educational and Psychosocial Measurement,
20, 1960, pp. 37-46.
[45] J.R. Landis, and G.G. Koch, “The measurement of
observer agreement for categorical data”, Biometrics, 1977, pp. 159-174.
[46] M. Greenacre, Correspondence Analysis in Practice
(2nd ed.), Chapman & Hall/CRC, Boca Raton, FL, 2007.
[47] J.J. Inman, V. Shankar, and R. Ferraro, “The roles of
channel-category associations and geodemographics
in channel patronage”, Journal of Marketing, 68, 2004, pp. 51-71.
[48] J. Rojas-Mendez, and M.J. Hine, “South American
countries’ positioning on personality traits: analysis of 10 national tourism websites”, Journal of Vacation
Marketing, 2016, Available online first at:
http://jvm.sagepub.com/cgi/reprint/1356766716649227v1.pdf?ijkey=JiURVSvkJcadRB0&keytype=finite
[49] R. Opoku, “Mapping destination personality in
cyberspace: An evaluation of country web sites using
correspondence analysis”, Journal of Internet Commerce, 8, 2009, pp. 70-87.
[50] L.F. Pitt, R. Opoku, M. Hultman, R. Abratt, and S.
Spyropoulou, “What I say about myself: Communication of brand personality by African
countries”, Tourism Management, 28(3), 2007, pp.
835-844. [51] H. Askell-Williams, and M.J. Lawson, “A
correspondence analysis of child-care students’ and
medical students’ knowledge about teaching and
learning”, International Education Journal, 5(2), 2004, pp. 176-204.
959